void mpir_fft_butterfly_twiddle(mp_ptr u, mp_ptr v,
mp_ptr s, mp_ptr t, mp_size_t limbs, mp_bitcnt_t b1, mp_bitcnt_t b2)
{
mp_limb_t nw = limbs*GMP_LIMB_BITS;
mp_size_t x, y;
int negate1 = 0;
int negate2 = 0;
if (b1 >= nw)
{
negate2 = 1;
b1 -= nw;
}
x = b1/GMP_LIMB_BITS;
b1 = b1%GMP_LIMB_BITS;
if (b2 >= nw)
{
negate1 = 1;
b2 -= nw;
}
y = b2/GMP_LIMB_BITS;
b2 = b2%GMP_LIMB_BITS;
mpir_butterfly_lshB(u, v, s, t, limbs, x, y);
mpn_mul_2expmod_2expp1(u, u, limbs, b1);
if (negate2) mpn_neg_n(u, u, limbs + 1);
mpn_mul_2expmod_2expp1(v, v, limbs, b2);
if (negate1) mpn_neg_n(v, v, limbs + 1);
}
void fft_adjust_sqrt2(mp_limb_t * r, mp_limb_t * i1,
mp_size_t i, mp_size_t limbs, mp_bitcnt_t w, mp_limb_t * temp)
{
mp_bitcnt_t wn = limbs*FLINT_BITS;
mp_limb_t cy;
mp_size_t j = i/2, k = w/2;
mp_size_t y;
mp_bitcnt_t b1;
int negate = 0;
b1 = j + wn/4 + i*k;
if (b1 >= wn)
{
negate = 1;
b1 -= wn;
}
y = b1/FLINT_BITS;
b1 = b1%FLINT_BITS;
/* multiply by 2^{j + wn/4 + i*k} */
if (y)
{
mpn_copyi(temp + y, i1, limbs - y);
cy = mpn_neg_n(temp, i1 + limbs - y, y);
temp[limbs] = 0;
mpn_addmod_2expp1_1(temp + y, limbs - y, -i1[limbs]);
mpn_sub_1(temp + y, temp + y, limbs - y + 1, cy);
mpn_mul_2expmod_2expp1(r, temp, limbs, b1);
} else
mpn_mul_2expmod_2expp1(r, i1, limbs, b1);
/* multiply by 2^{wn/2} */
y = limbs/2;
cy = 0;
mpn_copyi(temp + y, r, limbs - y);
temp[limbs] = 0;
if (y) cy = mpn_neg_n(temp, r + limbs - y, y);
mpn_addmod_2expp1_1(temp + y, limbs - y, -r[limbs]);
mpn_sub_1(temp + y, temp + y, limbs - y + 1, cy);
/* shift by an additional half limb (rare) */
if (limbs & 1)
mpn_mul_2expmod_2expp1(temp, temp, limbs, FLINT_BITS/2);
/* subtract */
if (negate)
mpn_sub_n(r, r, temp, limbs + 1);
else
mpn_sub_n(r, temp, r, limbs + 1);
}
void mpir_fft_adjust(mp_ptr r, mp_ptr i1, mp_size_t i, mp_size_t limbs, mp_bitcnt_t w)
{
mp_bitcnt_t b1;
mp_limb_t cy;
mp_size_t x;
b1 = i*w;
x = b1/GMP_LIMB_BITS;
b1 = b1%GMP_LIMB_BITS;
if (x)
{
mpn_copyi(r + x, i1, limbs - x);
r[limbs] = 0;
cy = mpn_neg_n(r, i1 + limbs - x, x);
mpn_addmod_2expp1_1(r + x, limbs - x, -i1[limbs]);
mpn_sub_1(r + x, r + x, limbs - x + 1, cy);
mpn_mul_2expmod_2expp1(r, r, limbs, b1);
} else
mpn_mul_2expmod_2expp1(r, i1, limbs, b1);
}
void fft_butterfly(mp_limb_t * s, mp_limb_t * t, mp_limb_t * i1,
mp_limb_t * i2, mp_size_t i, mp_size_t limbs, mp_bitcnt_t w)
{
mp_size_t y;
mp_bitcnt_t b1;
b1 = i*w;
y = b1/FLINT_BITS;
b1 = b1%FLINT_BITS;
butterfly_lshB(s, t, i1, i2, limbs, 0, y);
mpn_mul_2expmod_2expp1(t, t, limbs, b1);
}
